After a day with Jev: a blazing-fast classifier, not a GPT replacement

jiayuan_jy · x · 2026-09-18

After spending a day examining @typesafeai's Jev, the author concludes the novelty wears off quickly: it's essentially a faster, constrained general-purpose classifier/decision-maker that an LLM can already replicate.

Key takeaways:

The underlying model size is unknown (probably small), so its world knowledge may lag a typical LLM, raising reliability questions in complex situations
Best fit: tasks with a bounded set of choices and tight latency requirements; the constrained output format lets you guarantee valid outputs in code
Computer use is a natural fit: page DOM elements form a finite set, enabling a loop of "DOM list → Jev action → new DOM list"; but reasoning over long contexts is unproven — the author guesses it won't match GPT 6 Astra, just much faster
It can't replace tool use: tool call arguments (e.g., specific lines and replacement text for an edit tool) can't be enumerated upfront. Rewriting parts of Pi Agent with Jev yielded fewer opportunities than expected; compaction works but LLMs do it too
Bigger potential in decision-tree domains with demanding latency: game AI, robotics, autonomous driving

Experiments in progress: (1) a Poker AI with head-to-head battles between Jev and other models; (2) a Pokémon VGC (competitive doubles) AI — plenty of data, no good practice opponent exists yet, with plans to take it into ranked matches. Both are turn-based, so LLMs can handle them too, making them testbeds rather than clear wins for Jev.

Original post →

After a day with Jev: a blazing-fast classifier, not a GPT replacement

More from coding & agent